Xiaomi Brings Language Scaling to Robotics: Pre-trained on 100k Hours of Data

andrew_n_carr · x · 2026-07-22

Xiaomi has applied language model scaling laws to robotics. They pre-trained a robotics policy on 100,000 hours of egocentric data, followed by fine-tuning on 7,000 hours of robot trajectories.

The scaling gains show no signs of saturation: the real-robot success rate continues to improve with more pre-training data and larger models. This highlights the fundamental importance of massive data scale in embodied AI.

Original post →

More from Embodied

Embodied channel →